If you've ever felt like ChatGPT or Claude got noticeably worse a few weeks after a new version launched, you're not imagining it. A developer has built a dashboard that visualizes historical ELO ratings from Chatbot Arena - the crowd-sourced benchmark where real users vote on which AI response they prefer - to test exactly that hypothesis.
The tracker plots one continuous performance curve per flagship model over time, instead of the usual tangle of every minor variant and sub-version. That design choice makes a pattern visible: models often debut at a rating peak, then drift downward over the following weeks as a broader, more diverse user base tests them.
ELO is a scoring system borrowed from chess - a model's score rises when it beats a competitor in a head-to-head vote and falls when it loses. Chatbot Arena has run millions of these comparisons, making its leaderboard one of the more reliable real-world signals for model quality.
The post-launch drift has a few possible explanations: the early user pool skewing toward enthusiasts who favor particular response styles, providers quietly adjusting models through inference-side changes (modifications to how the model runs in production, not to the underlying model itself), or simple regression to the mean as hype fades. The dashboard doesn't answer which is correct, but it makes the pattern trackable over time instead of relying on gut feel.
For anyone who manages AI tool workflows or advises clients on model selection, historical trajectory is more useful than a snapshot leaderboard. A model's current rank tells you where it stands today; its six-week trend tells you whether it has been quietly declining since the launch-day press coverage.